Sketch recognition method and system based on multi-layer residual learning

Abstract

The invention discloses a sketch recognition method and system based on multilayer residual learning. The method comprises the following steps: acquiring a sketch sample, wherein the sketch sample comprises a training sample and a test sample; performing quantity enhancement processing on the sketch sample; constructing a multi-layer residual network; training the multi-layer residual network through the training sample to obtain a sketch recognition model; and identifying the sketch through the sketch training model. According to the method, the sketch is processed as a static graph, so that the sketch training set does not need to contain stroke sequence information during training, and the training input is a sample instead of a sample pair, so that the manual labeling cost is reduced; and the training method is simple, and the sketch recognition rate is high.

Description

technical field [0001] The invention relates to the field of sketch recognition, in particular to a sketch recognition method and system based on multi-layer residual learning. Background technique [0002] Unlike conventional images, sketches are a simple and efficient communication tool that can describe objects and express scenes or events that are difficult to describe with words. Therefore, the research on sketches has attracted more and more attention of researchers. Previous research results include handcrafted feature methods and deep learning methods. The former mainly continues the traditional technology of image recognition. Its idea is to extract manually designed artificial features and combine them with aggregation expression methods for local features (such as: bag-of-words model, etc.) to generate a final representation, which is then sent to a classifier for recognition. The latter may combine the characteristics of sketches such as: stroke timing informat...

AI Technical Summary

Problems solved by technology

Although the deep learning method has achieved good performance, as far as the present invention is concerned, most high-performance recognition methods need additional information or operations such as: timing information of hand-drawn sketch strokes, data enhancement operations, etc., which undoubtedly improves the performance of the network. Requirements for dependent datasets

In addition, many models have undergone secondary training in order to improve the ability of sketch recognition, which increases the training cost of the network

Images